2007-8 MBP with GeForce 8600M also was a widely known dud. Mine died, got a free mobo replacement in 2010 when Apple extended the warranty on these models, and it died again after a year.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
AMD Confirms RX 480 At $199 USD, Other APU & Polaris Announcements
- Thread starter lowendlinux
- Start date
- Sort by reaction score
Sadly the 2010 MBPs also had defective GPUs, NVIDIAs GT330M and GT320M which broke down over time. That was more hush hush, I never knew about the recall and missed the window on my 15" MBP.
For comparison my 4,1 to 5,1 with dual X5677 and Titan X. Impressive to see the 4 Skylake cores effectively matching my 8 core. I can't WAIT to build a modern PC!!!
http://www.3dmark.com/3dm/13266198
Thanks, this really shows how much those old CPUs are bottlenecking on Time Spy because the Titan X would get about 500 points higher with my Skylake. With one of the new 6-8 core Broadwells it would be even higher.
Still it's a good score for a 5-6 year old architecture paired with modern GPU.
Np! Yeah I am not complaining. It's true you see the growing hit any GPU is taking in cMP these days though, thank goodness for el cheapo craigslist GPUs 
Um, I just returned a RX480 and got a 1080 for this very reason (and obviously also the 5x performance gain and didn't want to crossfire RX480). nVidia drivers are much much better optimized in general.
Sure in day to day performance they're all similar now, but I prefer nVidia (personally) because they release far better drivers in a timely manner.
AMD only offers higher compute powers in OpenCL, which is why Apple keeps going to them. But in CUDA, the computational power is some of the best in the industry. Take Otoy's Octane renderer for example. With CUDA, the performance is astounding. It doesn't support OpenCL yet, though, but Otoy has an internal OpenCL version going for testing purposes and they state that CUDA just kills OpenCL in comparable cards.
Take other apps for example, like Adobe apps and Davinci Resolve, which both support CUDA. CUDA is far superior.
AMD is good, but does not have the raw graphics power that nVidia has, and that to me is important.
AMD is still a good company and it's good to have competition, but nVidia and AMD are very different companies.
As far as Apple's target for "Video editing" that's a misnomer. NO ONE in the industry uses FCPX. The film/pro industry ONLY uses Avid (No acceleration) or Premiere (Acceleration via OpenCL or CUDA).
So if you're talking about professional apps, Apple has no say in this at all. For Metal, it doesn't even matter anyway, since the GPU requirements are so low, any card will work and it doesn't need crazy amounts of computational power.
If you're data mining for bitcoin, sure go ahead and get an AMD GPU. Or if you're a consumer and using FCPX (Note no pro uses FCPX) go ahead and use an AMD GPU.
Compute is just that: mathemathical algorithms. The amount of power is defined by how much TFLOPs can a GPU push.
For example: R9 390X will be very close in terms of compute performance to GTX 980 TI, because both have similar compute power ~ 6 TFLOPs. CUDA, and OpenCL are just APIs that expose that for applications. There is nothing in CUDA that makes 6 TFLOPs Nvidia GPU faster than 6 TFLOPs GPU from AMD.
On the other hand applications using CUDA are better optimised for Nvidia hardware because that is proprietary Nvidia API, that is not available anywhere else.
Other hand of this is: Developers are not always competent enough to see that Application ported from CUDA might suffer.
There is word in the industry about people testing this internally, about the nature of CUDA, that when you port from CUDA designed application to OpenCL, you end up with lower performance on other GPUs(AMD GPUs suffer). This behaviour is not observed if you have the same application designed for OpenCL from the beginning, and ported for CUDA(GPUs from different architectures perform pretty much on the same level). But that remains to be confirmed.
The rest of your post is simple overgeneralising on the matter.
my 8600M died in my laptop and got replaced to, real pain. (still works but the screen died, run it with an external display for fun)
valve likes volcano so i gess there helping to push it https://en.wikipedia.org/wiki/Vulkan_(API) as it will help bring linux up to windows fps (or closer).
some things are relay pointless to talk about tho, if you need a app that runs on CUDA for work then you will need a Nvidia card that simple. you look at the apps you use and see what works best for you in your budget at the time you need it.
for most normal ppl CUDA is not a thing adobe is moving away from it, so it's only a few more nitch apps using it (big data centers or uni's doing big data crushing is not a normal user).
and still both the GTX10XX+the RX 480 do not work in osx so we do not know what will work best yet.. maybe nvida will dump OSX support or maybe they will actually relay move things on with well supported drivers (well supported is windows level updates to drivers not once an osx update). maybe the RX 480 will just run better in osx than the nvida cards (if we see drivers like the ones for the GTX 9XX cards in osx.
valve likes volcano so i gess there helping to push it https://en.wikipedia.org/wiki/Vulkan_(API) as it will help bring linux up to windows fps (or closer).
some things are relay pointless to talk about tho, if you need a app that runs on CUDA for work then you will need a Nvidia card that simple. you look at the apps you use and see what works best for you in your budget at the time you need it.
for most normal ppl CUDA is not a thing adobe is moving away from it, so it's only a few more nitch apps using it (big data centers or uni's doing big data crushing is not a normal user).
and still both the GTX10XX+the RX 480 do not work in osx so we do not know what will work best yet.. maybe nvida will dump OSX support or maybe they will actually relay move things on with well supported drivers (well supported is windows level updates to drivers not once an osx update). maybe the RX 480 will just run better in osx than the nvida cards (if we see drivers like the ones for the GTX 9XX cards in osx.
What is, that nVidia makes better GPU's than AMD?
It's not an opinion, it's a fact.
Like I said, AMD is great in it's own right for affordable GPU's and computational OpenCL operations.
I like AMD, and purchased an RX480 (with plans to crossfire) but after seeing the 1080/1070 performance, I jumped on the 1080 because I need to use VR for dev purposes.
Remember, TFLOPS only refers to theoretical performance. I'll assume by power you mean performance. Performance can only be determined by determining how the GPU performs on certain tasks. Sure, the 390X had very similar TFLOPS to the GTX 980 Ti, but in gaming tasks and single precision compute tasks the 980 Ti performed much better than the 390X.
I'm not going to wade into the debate of whether OpenCL or CUDA is better. For whats it worth, my experience in academia has shown that CUDA has much broader support and even engineering classes are taught in how to program in CUDA. However I get why Apple would stick to OpenCL and Metal. Its the only library they can run across all macs (including those without discrete GPUs) and it doesn't tie them to a single GPU vendor.
Given Nvidia's lead in GPU efficiency the last couple generations its been disappointing that they haven't ended up in any macs. Especially when Apple only uses GPUs that are relatively low power.
I've tried to highlight the risks in blindly accepting theoretical TFLOPS as deliverable performance, but to no effect.
Who cares how many TFLOPS the 2nd GPU on an MP6,1 could theoretically produce - when most of the time in most of the applications it's sitting idle and wasting power?
Similarly, it's pointless to debate the ratio of FP32 to FP64 if your application doesn't need FP64. Nvidia knows this, and has the Kepler and GP100 for FP64 tasks, and went with weak FP64 on the Maxwell.
Last edited:
I think we need to define "better GPU", otherwise hard to tell what we are debating on. Is it really a fact? Or opinion.
If only talking about performance, of course the real world performance is much much more important to end user. However, IMO, that's a better "overall experence" which combine both hardware and software. It's really hard to tell if Nvidia really makes "better" GPU then AMD.
If real world exp is everything, then may be AMD make better GPU when using FCPX. It's so hard to fine a "fair benchmark" to compare the GPUs in all aspect. If AMD's GPU is more powerful in OpenCL and Nvidia's GPU is more powerful in OpenGL, then which one is more powerful? That's hard to compare.
I think there is no doubt that Nvidia gives the end user better exp most of them time, but I personally still can't conclude that Nvidia makes better GPU (in terms pure hardware).
To make it extreme and easier to understand. If GPU A is faster then GPU B by 100% in all benchmark. But GPU B crash rate is 100% higher then GPU A (assume the crash rate is not zero). Then which GPU is better? No matter which one we choose, we are choosing the better experence base on our own opinion, right?
Or GPU C has better architecture, better protential, however very poor software / driver support, so only 50% of its power can be used. GPU D's protential Max performance is only 80% of GPU C, but the software make it able to deliver 100% performance. In this case, which GPU (pure hardware) is better? In terms of user exp, of course GPU D, but GPU C is stronger in hardware, just waiting for software support, can it actually be the better GPU?
If we don't have the same definition, it's hard to get a mutual agreed result.
What Koyoot focus on is the max protential, I agree that's basically means nothing to the end user, however, I cannot deny that may be a way to define "better GPU".
If we keep debating which GPU is better but no clear definition. That is a bit like comparing and i5 to a Xeon. The i5 has less core, better single core performance but only 50% multi core performance of the Xeon does. The Xeon has much more cores, but much slower single core performance. So now, which CPU is better? Or Stronger? I think it's very clear that depends on the job, right?
Um, it's a fact. Heh.
It's like comparing a Toyota Camry (AMD) to a Lamborghini (nVidia). It's not a fair comparison.
RX480 8GB is $250. GTX1070 is $450. GTX1080 is $600.
Look at the technical specs for all these cards. And then look at the price. No one is saying the RX480 is not a good value, it actually is a very good value for the price. But it's a value card, a value car, a Toyota Camry.
The only reason FCPX has good performance with AMD is because Apple tuned it for the AMD cards. Actually, FCPX does better on old 280X cards than on something like a 390X.
Fact is, no one important uses FCPX so it doesn't really matter. ANYONE who does Workstation level work switched to HP or Dell workstations with Xeon's and Workstation GPUs.
Fact is, EVEN if OpenCL computational is a standard, and AMD is "better" at it doesn't really matter. By the time Apple puts in a new GPU in their Mac, it's old already within 6 months. The FireGL cards in the Mac Pros are old as hell, so what does it matter if it has better OpenCL capability? The GPU market moves faster than the CPU market.
nVidia literally murdered the whole industry with the 16nm FINFET GPU. It's low power, high performance is incredible.
2x Crossfire RX480's don't even reach a single GTX1080 performance, and plus they eat almost twice as much power as a single GTX1080.
I'd be surprised if, and when, the nVidia drivers for 1080 are released for OS X, that the 1080 doesn't have good OpenCL performance. I'd be first to check here and post results.
Anyway, this is an AMD thread I'll stay out of it, but my point still stands, AMD is a good company and they have other ideals, they compete in other areas and aren't "gamegeeks" like nVidia. I just wish Apple would give us the option for BOTH AMD or nVidia GPUs during ordering for their Macs.
Hopefully, IF eGPU via TB3 is "officially" supported by Apple, drivers/support will come to us much faster for both AMD and nVidia.
Apple really needs to rethink their Mac Pro strategy, and maybe step away from using workstation hardware.
Last edited:
Pascal even weaker at FP64. Don't know how well GP100 will perform compared to previous gens or competition.
Simulations and modelling by industry industry and science are becoming more dependent on servers housing thousands of CPUs. Distributed computing is going online. The latest After Effects betas even have an option to use Google's server to rent CPU time.
So perhaps having FP64 performance at the desktop level isn't important anymore. Nvidia would prefer people to use GRID. AMD has no server based solutions here so workstation performance is their priority.
Guys, the only reason why particular hardware can be idling in applications is... Applications themselves. CUDA has this advantage over OpenCL that you do not need to optimise your applications for it. Thats what makes difference! It from the start allows perfect 100% load on the GPU and 100% usage of features of the hardware. On OpenCL you HAVE TO optimise the application, for hardware features to get it work 100%. With OpenCL feature set you can make general application, like Luxmark benchmark, that is not gimping anyones performance anywhere, unless you will optimise it for specific hardware that will block optimisations for others. Im looking at green brand here!
I have said this before, I will repeat this: Ask developers to not be bloody lazy! Make good applications, that utilize 100% of GPU capabilities. Like for example Final Cut Pro X that is properly coded for OpenCL.
Especially when you say about gaming tasks. Vulkan, DX12 games show that properly coded application will make 6 TFLOPS GPU from AMD(R9 390X) perform on the same level as 6 TFLOPs GPU from Nvidia(GTX 980 Ti). For example here: https://www.computerbase.de/2016-07...md-nvidia/#diagramm-doom-mit-vulkan-2560-1440
Look how close to GTX 980 Ti which should have 6 TFLOPs is 5.1 TFLOPs GPU(R9 390). And how far away from GTX 1070 is Fury X. Vulkan did not made AMD GPUs perform better. Nor did DX12. It just made them work as they should from the ground up without software bottlenecks (high CPU overhead).
Applications are using hardware. Remember that and start asking developers for properly coded software.
I have said this before, I will repeat this: Ask developers to not be bloody lazy! Make good applications, that utilize 100% of GPU capabilities. Like for example Final Cut Pro X that is properly coded for OpenCL.
It does not matter which one is better. None of is better than another. CUDA does not make your GPUs faster just because. It is heavily optimised for Nvidia hardware, and your devs do not have to do anything when they implement it into their applications. But perception in the world is simple: Nvidia makes better GPUs. Even if they don't.
Especially when you say about gaming tasks. Vulkan, DX12 games show that properly coded application will make 6 TFLOPS GPU from AMD(R9 390X) perform on the same level as 6 TFLOPs GPU from Nvidia(GTX 980 Ti). For example here: https://www.computerbase.de/2016-07...md-nvidia/#diagramm-doom-mit-vulkan-2560-1440
Look how close to GTX 980 Ti which should have 6 TFLOPs is 5.1 TFLOPs GPU(R9 390). And how far away from GTX 1070 is Fury X. Vulkan did not made AMD GPUs perform better. Nor did DX12. It just made them work as they should from the ground up without software bottlenecks (high CPU overhead).
Nvidia does not make better GPU's. Fiji XT made on 28nm process is still faster in compute applications, that are open and Multiplatform than GTX 1080, which has been proven in this thread, recently. Without understanding big scheme your perception of situation will be faulty, unfortunately.
Applications are using hardware. Remember that and start asking developers for properly coded software.
Forget about this benchmark, as something meaningful for DX12. One guy from Futuremark company came to Anandtech Forum and says that... It is DX12 FL11_0. What this means is that there is no most important parts of DX12 in the rendering pipeline, including... Asynchronous Compute.
Last edited:
It's a good stress test even if using a more backward compatible feature set. However, there are plenty of front page articles showing Async on and off. Gains are seen on Radeons and Pascals. Maxwell nothing, even a loss.
Thats because Time Spy is NOT using Asynchronous Compute Shaders, but simple preemption. Big difference.
Nvidia uses SOFTWARE technique of pre-emptying the calls. AMD uses HARDWARE level Asynchronous Compute - running Compute AND graphics in parallel.
Nothing to see here really, with that benchmark. Truth be told, only benchmarks that show performance of DX12, and Vulkan and VR applications on GPUs currently are... DX12, Vulkan, VR games. And there are not that many examples, however there is few titles coming up next...: Battlefield 1, Civ 6, Deus Ex.
Also about Vega architecture.
Today a word happened to say that Vega may have new scheduler that is graphics and compute agnostic. Previous generations of GCN would have 8 ACE(Asynchronous Compute Engines) and 1 graphics Command Scheduler. Later implementation like in Fiji have had 4 ACE's, 1 Graphics Command Sceduler and 2 HWS(Hardware Schedulers). You can see where AMD might want to go from here. Both compute and graphics in SINGLE command scheduler. What is interesting here is that Mantle was designed for it from ground up, and Vulkan and Metal are doing exactly that - combine OpenGL(Graphics) and OpenCL(Compute) into single queue. This is my understanding from the word that next gen. scheduler can combine both capabilities into one, it is not anything official. But few things are logical, at least by looking at other AMD offerings.
Time will tell if this will happen to be true, and if it will happen - how it will affect performance. However, this is the front on which AMD architecture was lacking - graphics scheduling.
Edit: after thinking for a moment about this possibility, it would make on hardware level for AMD what Nvidia is doing on software level with CUDA. Extremely interesting...
Last edited:
Pascal is not weaker than maxwell. Check out this benchmark from anandtech. Nvidia diverged their gaming GPUs and their compute GPUs when they released maxwell. Nvidia has the luxury of being able to design a lot more chips that fit in smaller niches. Thats why they have GP100, a compute focused GPU with a monstrous 5 TFLOPS of FP64 compute performance and GP104 with its ~9 TFLOPS of single precision compute performance. I am using TFLOPS here because it allows us to ballpark the relative performance.
This is naive. These low level APIs like metal, directx12 and vulkan give the programmer more control, but it also requires more work. If you are a manager at a software company, you may very well say its not worth days/weeks/months of optimization to get out a small increase in performance. Especially if that performance increase only applies to a subset of all your users.
Additionally, just because you use these APIs does not automatically gain you more performance. The most recent version of tomb raider came out with a directx12 code path and saw no performance benefit on either AMD or Nvidia hardware.
Again, this depends on the application you are looking at. Even if we disregard CUDA, Nvidia is still faster in some single precision compute benchmarks. Look here, here, here and here. Fiji had a 1:16 ratio of SP/DP compute units, meaning that it still gets beat by Hawaii in DP tasks.
In one paragraph of your post you state something for it to in some way contradict in another.
.I know perfectly well those benchmarks, nor I say that AMD is solution for everything, or say that Nvidia is solution for everything. We have been discussing OpenCL compute rendering on previous pages, and differences in OCL benchmarks with Asgorath. AMD is better in compute rendering, Nvidia in image analysing.
About Applications, and optimisation. Drivers and optimising them for applications like gaming let AMD bring up to 50% of performance from their hardware over past year(this is especially true for Grenada/Hawaii chips). So those are not small differences. In a perfect world, every application would be tuned with time for hardware, but in most cases devs are getting over themselves and simply... Ignore that part, when 90% of application is running stable. It is possible to understand. But saying that one hardware is better than other because Devs are lazy and do not want to optimise the software for hardware is...
.But I suppose that is how marketing and mindshare works.
Look at the clockspeed. Per clock Pascal is down from Maxwell. Per clock Maxwell was down from Kepler.
So what? GPUs don't need to compete in performance per clock. Nvidia wanted to take advantage of the high clocks they could achieve with the 16 nm process. In every performance and efficiency metric Pascal beats Maxwell hands down.
How come in OpenCL GTX 980 was faster than GTX 780 Ti, even if Maxwell had lower compute power than GTX 780 Ti?
Because GTX 980 had much improved compute capabilities compared to Kepler
.Also in not every GPU rendering benchmark GTX 1080 is slower than previous gen.
https://www.computerbase.de/2016-06...5/#abschnitt_gpucomputing_und_videowiedergabe
I have had Kepler, Maxwell, Pascal. I'm not going to blind myself from accepting that Nvidia's performance per clock is inefficient and that they are trying to play a megahertz war against AMD.
It's good that they can clock up speeds and reduce power consumption, but let's not lose sight of the fact that this horse was beaten to death during the Intel vs AMD megahertz war when Intel tried the same gimmick with the Pentium 4. Very high clock speed but **** performance clock for clock against their own older gens.
[doublepost=1468862914][/doublepost]
How come in OpenCL GTX 980 was faster than GTX 780 Ti, even if Maxwell had lower compute power than GTX 780 Ti?
How about adding the Titan or Titan Black?